src/gallium/auxiliary/gallivm/lp_bld_misc.cpp

   1 /**************************************************************************
   2  *
   3  * Copyright 2010 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  18  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  19  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  20  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * The above copyright notice and this permission notice (including the
  23  * next paragraph) shall be included in all copies or substantial portions
  24  * of the Software.
  25  *
  26  **************************************************************************/
  27
  28
  29 /**
  30  * The purpose of this module is to expose LLVM functionality not available
  31  * through the C++ bindings.
  32  */
  33
  34
  35 // Undef these vars just to silence warnings
  36 #undef PACKAGE_BUGREPORT
  37 #undef PACKAGE_NAME
  38 #undef PACKAGE_STRING
  39 #undef PACKAGE_TARNAME
  40 #undef PACKAGE_VERSION
  41
  42
  43 #include <stddef.h>
  44
  45 #include <llvm/Config/llvm-config.h>
  46
  47 #if LLVM_VERSION_MAJOR < 7
  48 // Workaround http://llvm.org/PR23628
  49 #pragma push_macro("DEBUG")
  50 #undef DEBUG
  51 #endif
  52
  53 #include <llvm/Config/llvm-config.h>
  54 #include <llvm-c/Core.h>
  55 #include <llvm-c/Support.h>
  56 #include <llvm-c/ExecutionEngine.h>
  57 #include <llvm/Target/TargetOptions.h>
  58 #include <llvm/ExecutionEngine/ExecutionEngine.h>
  59 #include <llvm/ADT/Triple.h>
  60 #include <llvm/Analysis/TargetLibraryInfo.h>
  61 #include <llvm/ExecutionEngine/SectionMemoryManager.h>
  62 #include <llvm/Support/CommandLine.h>
  63 #include <llvm/Support/Host.h>
  64 #include <llvm/Support/PrettyStackTrace.h>
  65
  66 #include <llvm/Support/TargetSelect.h>
  67
  68 #include <llvm/IR/CallSite.h>
  69 #include <llvm/IR/IRBuilder.h>
  70 #include <llvm/IR/Module.h>
  71 #include <llvm/Support/CBindingWrapping.h>
  72
  73 #include <llvm/Config/llvm-config.h>
  74 #if LLVM_USE_INTEL_JITEVENTS
  75 #include <llvm/ExecutionEngine/JITEventListener.h>
  76 #endif
  77
  78 #if LLVM_VERSION_MAJOR < 7
  79 // Workaround http://llvm.org/PR23628
  80 #pragma pop_macro("DEBUG")
  81 #endif
  82
  83 #include "c11/threads.h"
  84 #include "os/os_thread.h"
  85 #include "pipe/p_config.h"
  86 #include "util/u_debug.h"
  87 #include "util/u_cpu_detect.h"
  88
  89 #include "lp_bld_misc.h"
  90 #include "lp_bld_debug.h"
  91
  92 namespace {
  93
  94 class LLVMEnsureMultithreaded {
  95 public:
  96    LLVMEnsureMultithreaded()
  97    {
  98       if (!LLVMIsMultithreaded()) {
  99          LLVMStartMultithreaded();
 100       }
 101    }
 102 };
 103
 104 static LLVMEnsureMultithreaded lLVMEnsureMultithreaded;
 105
 106 }
 107
 108 static once_flag init_native_targets_once_flag = ONCE_FLAG_INIT;
 109
 110 static void init_native_targets()
 111 {
 112    // If we have a native target, initialize it to ensure it is linked in and
 113    // usable by the JIT.
 114    llvm::InitializeNativeTarget();
 115
 116    llvm::InitializeNativeTargetAsmPrinter();
 117
 118    llvm::InitializeNativeTargetDisassembler();
 119 #if DEBUG
 120    {
 121       char *env_llc_options = getenv("GALLIVM_LLC_OPTIONS");
 122       if (env_llc_options) {
 123          char *option;
 124          char *options[64] = {(char *) "llc"};      // Warning without cast
 125          int   n;
 126          for (n = 0, option = strtok(env_llc_options, " "); option; n++, option = strtok(NULL, " ")) {
 127             options[n + 1] = option;
 128          }
 129          if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
 130             debug_printf("llc additional options (%d):\n", n);
 131             for (int i = 1; i <= n; i++)
 132                debug_printf("\t%s\n", options[i]);
 133             debug_printf("\n");
 134          }
 135          LLVMParseCommandLineOptions(n + 1, options, NULL);
 136       }
 137    }
 138 #endif
 139 }
 140
 141 extern "C" void
 142 lp_set_target_options(void)
 143 {
 144    /* The llvm target registry is not thread-safe, so drivers and state-trackers
 145     * that want to initialize targets should use the lp_set_target_options()
 146     * function to safely initialize targets.
 147     *
 148     * LLVM targets should be initialized before the driver or state-tracker tries
 149     * to access the registry.
 150     */
 151    call_once(&init_native_targets_once_flag, init_native_targets);
 152 }
 153
 154 extern "C"
 155 LLVMTargetLibraryInfoRef
 156 gallivm_create_target_library_info(const char *triple)
 157 {
 158    return reinterpret_cast<LLVMTargetLibraryInfoRef>(
 159    new llvm::TargetLibraryInfoImpl(
 160    llvm::Triple(triple)));
 161 }
 162
 163 extern "C"
 164 void
 165 gallivm_dispose_target_library_info(LLVMTargetLibraryInfoRef library_info)
 166 {
 167    delete reinterpret_cast<
 168    llvm::TargetLibraryInfoImpl
 169    *>(library_info);
 170 }
 171
 172
 173 typedef llvm::RTDyldMemoryManager BaseMemoryManager;
 174
 175
 176 /*
 177  * Delegating is tedious but the default manager class is hidden in an
 178  * anonymous namespace in LLVM, so we cannot just derive from it to change
 179  * its behavior.
 180  */
 181 class DelegatingJITMemoryManager : public BaseMemoryManager {
 182
 183    protected:
 184       virtual BaseMemoryManager *mgr() const = 0;
 185
 186    public:
 187       /*
 188        * From RTDyldMemoryManager
 189        */
 190       virtual uint8_t *allocateCodeSection(uintptr_t Size,
 191                                            unsigned Alignment,
 192                                            unsigned SectionID,
 193                                            llvm::StringRef SectionName) {
 194          return mgr()->allocateCodeSection(Size, Alignment, SectionID,
 195                                            SectionName);
 196       }
 197       virtual uint8_t *allocateDataSection(uintptr_t Size,
 198                                            unsigned Alignment,
 199                                            unsigned SectionID,
 200                                            llvm::StringRef SectionName,
 201                                            bool IsReadOnly) {
 202          return mgr()->allocateDataSection(Size, Alignment, SectionID,
 203                                            SectionName,
 204                                            IsReadOnly);
 205       }
 206       virtual void registerEHFrames(uint8_t *Addr, uint64_t LoadAddr, size_t Size) {
 207          mgr()->registerEHFrames(Addr, LoadAddr, Size);
 208       }
 209 #if LLVM_VERSION_MAJOR >= 5
 210       virtual void deregisterEHFrames() {
 211          mgr()->deregisterEHFrames();
 212       }
 213 #else
 214       virtual void deregisterEHFrames(uint8_t *Addr, uint64_t LoadAddr, size_t Size) {
 215          mgr()->deregisterEHFrames(Addr, LoadAddr, Size);
 216       }
 217 #endif
 218       virtual void *getPointerToNamedFunction(const std::string &Name,
 219                                               bool AbortOnFailure=true) {
 220          return mgr()->getPointerToNamedFunction(Name, AbortOnFailure);
 221       }
 222       virtual bool finalizeMemory(std::string *ErrMsg = 0) {
 223          return mgr()->finalizeMemory(ErrMsg);
 224       }
 225 };
 226
 227
 228 /*
 229  * Delegate memory management to one shared manager for more efficient use
 230  * of memory than creating a separate pool for each LLVM engine.
 231  * Keep generated code until freeGeneratedCode() is called, instead of when
 232  * memory manager is destroyed, which happens during engine destruction.
 233  * This allows additional memory savings as we don't have to keep the engine
 234  * around in order to use the code.
 235  * All methods are delegated to the shared manager except destruction and
 236  * deallocating code.  For the latter we just remember what needs to be
 237  * deallocated later.  The shared manager is deleted once it is empty.
 238  */
 239 class ShaderMemoryManager : public DelegatingJITMemoryManager {
 240
 241    BaseMemoryManager *TheMM;
 242
 243    struct GeneratedCode {
 244       typedef std::vector<void *> Vec;
 245       Vec FunctionBody, ExceptionTable;
 246       BaseMemoryManager *TheMM;
 247
 248       GeneratedCode(BaseMemoryManager *MM) {
 249          TheMM = MM;
 250       }
 251
 252       ~GeneratedCode() {
 253       }
 254    };
 255
 256    GeneratedCode *code;
 257
 258    BaseMemoryManager *mgr() const {
 259       return TheMM;
 260    }
 261
 262    public:
 263
 264       ShaderMemoryManager(BaseMemoryManager* MM) {
 265          TheMM = MM;
 266          code = new GeneratedCode(MM);
 267       }
 268
 269       virtual ~ShaderMemoryManager() {
 270          /*
 271           * 'code' is purposely not deleted.  It is the user's responsibility
 272           * to call getGeneratedCode() and freeGeneratedCode().
 273           */
 274       }
 275
 276       struct lp_generated_code *getGeneratedCode() {
 277          return (struct lp_generated_code *) code;
 278       }
 279
 280       static void freeGeneratedCode(struct lp_generated_code *code) {
 281          delete (GeneratedCode *) code;
 282       }
 283
 284       virtual void deallocateFunctionBody(void *Body) {
 285          // remember for later deallocation
 286          code->FunctionBody.push_back(Body);
 287       }
 288 };
 289
 290
 291 /**
 292  * Same as LLVMCreateJITCompilerForModule, but:
 293  * - allows using MCJIT and enabling AVX feature where available.
 294  * - set target options
 295  *
 296  * See also:
 297  * - llvm/lib/ExecutionEngine/ExecutionEngineBindings.cpp
 298  * - llvm/tools/lli/lli.cpp
 299  * - http://markmail.org/message/ttkuhvgj4cxxy2on#query:+page:1+mid:aju2dggerju3ivd3+state:results
 300  */
 301 extern "C"
 302 LLVMBool
 303 lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
 304                                         lp_generated_code **OutCode,
 305                                         LLVMModuleRef M,
 306                                         LLVMMCJITMemoryManagerRef CMM,
 307                                         unsigned OptLevel,
 308                                         char **OutError)
 309 {
 310    using namespace llvm;
 311
 312    std::string Error;
 313    EngineBuilder builder(std::unique_ptr<Module>(unwrap(M)));
 314
 315    /**
 316     * LLVM 3.1+ haven't more "extern unsigned llvm::StackAlignmentOverride" and
 317     * friends for configuring code generation options, like stack alignment.
 318     */
 319    TargetOptions options;
 320 #if defined(PIPE_ARCH_X86)
 321    options.StackAlignmentOverride = 4;
 322 #endif
 323
 324    builder.setEngineKind(EngineKind::JIT)
 325           .setErrorStr(&Error)
 326           .setTargetOptions(options)
 327           .setOptLevel((CodeGenOpt::Level)OptLevel);
 328
 329 #ifdef _WIN32
 330     /*
 331      * MCJIT works on Windows, but currently only through ELF object format.
 332      *
 333      * XXX: We could use `LLVM_HOST_TRIPLE "-elf"` but LLVM_HOST_TRIPLE has
 334      * different strings for MinGW/MSVC, so better play it safe and be
 335      * explicit.
 336      */
 337 #  ifdef _WIN64
 338     LLVMSetTarget(M, "x86_64-pc-win32-elf");
 339 #  else
 340     LLVMSetTarget(M, "i686-pc-win32-elf");
 341 #  endif
 342 #endif
 343
 344    llvm::SmallVector<std::string, 16> MAttrs;
 345
 346 #if LLVM_VERSION_MAJOR >= 4 && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64) || defined(PIPE_ARCH_ARM))
 347    /* llvm-3.3+ implements sys::getHostCPUFeatures for Arm
 348     * and llvm-3.7+ for x86, which allows us to enable/disable
 349     * code generation based on the results of cpuid on these
 350     * architectures.
 351     */
 352    llvm::StringMap<bool> features;
 353    llvm::sys::getHostCPUFeatures(features);
 354
 355    for (StringMapIterator<bool> f = features.begin();
 356         f != features.end();
 357         ++f) {
 358       MAttrs.push_back(((*f).second ? "+" : "-") + (*f).first().str());
 359    }
 360 #elif defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
 361    /*
 362     * We need to unset attributes because sometimes LLVM mistakenly assumes
 363     * certain features are present given the processor name.
 364     *
 365     * https://bugs.freedesktop.org/show_bug.cgi?id=92214
 366     * http://llvm.org/PR25021
 367     * http://llvm.org/PR19429
 368     * http://llvm.org/PR16721
 369     */
 370    MAttrs.push_back(util_cpu_caps.has_sse    ? "+sse"    : "-sse"   );
 371    MAttrs.push_back(util_cpu_caps.has_sse2   ? "+sse2"   : "-sse2"  );
 372    MAttrs.push_back(util_cpu_caps.has_sse3   ? "+sse3"   : "-sse3"  );
 373    MAttrs.push_back(util_cpu_caps.has_ssse3  ? "+ssse3"  : "-ssse3" );
 374    MAttrs.push_back(util_cpu_caps.has_sse4_1 ? "+sse4.1" : "-sse4.1");
 375    MAttrs.push_back(util_cpu_caps.has_sse4_2 ? "+sse4.2" : "-sse4.2");
 376    /*
 377     * AVX feature is not automatically detected from CPUID by the X86 target
 378     * yet, because the old (yet default) JIT engine is not capable of
 379     * emitting the opcodes. On newer llvm versions it is and at least some
 380     * versions (tested with 3.3) will emit avx opcodes without this anyway.
 381     */
 382    MAttrs.push_back(util_cpu_caps.has_avx  ? "+avx"  : "-avx");
 383    MAttrs.push_back(util_cpu_caps.has_f16c ? "+f16c" : "-f16c");
 384    MAttrs.push_back(util_cpu_caps.has_fma  ? "+fma"  : "-fma");
 385    MAttrs.push_back(util_cpu_caps.has_avx2 ? "+avx2" : "-avx2");
 386    /* disable avx512 and all subvariants */
 387    MAttrs.push_back("-avx512cd");
 388    MAttrs.push_back("-avx512er");
 389    MAttrs.push_back("-avx512f");
 390    MAttrs.push_back("-avx512pf");
 391    MAttrs.push_back("-avx512bw");
 392    MAttrs.push_back("-avx512dq");
 393    MAttrs.push_back("-avx512vl");
 394 #endif
 395 #if defined(PIPE_ARCH_ARM)
 396    if (!util_cpu_caps.has_neon) {
 397       MAttrs.push_back("-neon");
 398       MAttrs.push_back("-crypto");
 399       MAttrs.push_back("-vfp2");
 400    }
 401 #endif
 402
 403 #if defined(PIPE_ARCH_PPC)
 404    MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec");
 405 #if (LLVM_VERSION_MAJOR < 4)
 406    /*
 407     * Make sure VSX instructions are disabled
 408     * See LLVM bugs:
 409     * https://llvm.org/bugs/show_bug.cgi?id=25503#c7 (fixed in 3.8.1)
 410     * https://llvm.org/bugs/show_bug.cgi?id=26775 (fixed in 3.8.1)
 411     * https://llvm.org/bugs/show_bug.cgi?id=33531 (fixed in 4.0)
 412     * https://llvm.org/bugs/show_bug.cgi?id=34647 (llc performance on certain unusual shader IR; intro'd in 4.0, pending as of 5.0)
 413     */
 414    if (util_cpu_caps.has_altivec) {
 415       MAttrs.push_back("-vsx");
 416    }
 417 #else
 418    /*
 419     * Bug 25503 is fixed, by the same fix that fixed
 420     * bug 26775, in versions of LLVM later than 3.8 (starting with 3.8.1).
 421     * BZ 33531 actually comprises more than one bug, all of
 422     * which are fixed in LLVM 4.0.
 423     *
 424     * With LLVM 4.0 or higher:
 425     * Make sure VSX instructions are ENABLED (if supported), unless
 426     * VSX instructions are explicitly enabled/disabled via GALLIVM_VSX=1 or 0.
 427     */
 428    if (util_cpu_caps.has_altivec) {
 429       MAttrs.push_back(util_cpu_caps.has_vsx ? "+vsx" : "-vsx");
 430    }
 431 #endif
 432 #endif
 433
 434    builder.setMAttrs(MAttrs);
 435
 436    if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
 437       int n = MAttrs.size();
 438       if (n > 0) {
 439          debug_printf("llc -mattr option(s): ");
 440          for (int i = 0; i < n; i++)
 441             debug_printf("%s%s", MAttrs[i].c_str(), (i < n - 1) ? "," : "");
 442          debug_printf("\n");
 443       }
 444    }
 445
 446    StringRef MCPU = llvm::sys::getHostCPUName();
 447    /*
 448     * The cpu bits are no longer set automatically, so need to set mcpu manually.
 449     * Note that the MAttrs set above will be sort of ignored (since we should
 450     * not set any which would not be set by specifying the cpu anyway).
 451     * It ought to be safe though since getHostCPUName() should include bits
 452     * not only from the cpu but environment as well (for instance if it's safe
 453     * to use avx instructions which need OS support). According to
 454     * http://llvm.org/bugs/show_bug.cgi?id=19429 however if I understand this
 455     * right it may be necessary to specify older cpu (or disable mattrs) though
 456     * when not using MCJIT so no instructions are generated which the old JIT
 457     * can't handle. Not entirely sure if we really need to do anything yet.
 458     */
 459 #if PIPE_ARCH_LITTLE_ENDIAN  && defined(PIPE_ARCH_PPC_64)
 460    /*
 461     * Versions of LLVM prior to 4.0 lacked a table entry for "POWER8NVL",
 462     * resulting in (big-endian) "generic" being returned on
 463     * little-endian Power8NVL systems.  The result was that code that
 464     * attempted to load the least significant 32 bits of a 64-bit quantity
 465     * from memory loaded the wrong half.  This resulted in failures in some
 466     * Piglit tests, e.g.
 467     * .../arb_gpu_shader_fp64/execution/conversion/frag-conversion-explicit-double-uint
 468     */
 469    if (MCPU == "generic")
 470       MCPU = "pwr8";
 471 #endif
 472    builder.setMCPU(MCPU);
 473    if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
 474       debug_printf("llc -mcpu option: %s\n", MCPU.str().c_str());
 475    }
 476
 477    ShaderMemoryManager *MM = NULL;
 478    BaseMemoryManager* JMM = reinterpret_cast<BaseMemoryManager*>(CMM);
 479    MM = new ShaderMemoryManager(JMM);
 480    *OutCode = MM->getGeneratedCode();
 481
 482    builder.setMCJITMemoryManager(std::unique_ptr<RTDyldMemoryManager>(MM));
 483    MM = NULL; // ownership taken by std::unique_ptr
 484
 485    ExecutionEngine *JIT;
 486
 487    JIT = builder.create();
 488 #if LLVM_USE_INTEL_JITEVENTS
 489    JITEventListener *JEL = JITEventListener::createIntelJITEventListener();
 490    JIT->RegisterJITEventListener(JEL);
 491 #endif
 492    if (JIT) {
 493       *OutJIT = wrap(JIT);
 494       return 0;
 495    }
 496    lp_free_generated_code(*OutCode);
 497    *OutCode = 0;
 498    delete MM;
 499    *OutError = strdup(Error.c_str());
 500    return 1;
 501 }
 502
 503
 504 extern "C"
 505 void
 506 lp_free_generated_code(struct lp_generated_code *code)
 507 {
 508    ShaderMemoryManager::freeGeneratedCode(code);
 509 }
 510
 511 extern "C"
 512 LLVMMCJITMemoryManagerRef
 513 lp_get_default_memory_manager()
 514 {
 515    BaseMemoryManager *mm;
 516    mm = new llvm::SectionMemoryManager();
 517    return reinterpret_cast<LLVMMCJITMemoryManagerRef>(mm);
 518 }
 519
 520 extern "C"
 521 void
 522 lp_free_memory_manager(LLVMMCJITMemoryManagerRef memorymgr)
 523 {
 524    delete reinterpret_cast<BaseMemoryManager*>(memorymgr);
 525 }
 526
 527 extern "C" LLVMValueRef
 528 lp_get_called_value(LLVMValueRef call)
 529 {
 530         return LLVMGetCalledValue(call);
 531 }
 532
 533 extern "C" bool
 534 lp_is_function(LLVMValueRef v)
 535 {
 536         return LLVMGetValueKind(v) == LLVMFunctionValueKind;
 537 }